As methods of estimating a self-position at the time of image capture based on an image captured by a camera, methods (Patent Documents 1 and 2) of estimating a self-position from a plurality of landmarks detected from an image and a method (Non-Patent Document 1) of performing matching with a previously captured image group to which a position and a direction are assigned have been suggested. Here, the self-position refers to information indicating a position and an image capture direction at the time of image capture.
Patent Document 1 discloses a technology for estimating a self-position based on position information, in which a building, a mark, or the like used as a landmark is positioned in advance by a highly accurate GPS (Global Positioning System) device, and a camera parameter of an image capture device.
In the estimation method, an angle is calculated per pixel from a camera parameter of the camera used to capture an image, a constrained circle in which there is a probability that the self-position is present is defined from the maximum angle of nip and the minimum angle of nip between landmarks, and a score is calculated using a model which defines a likelihood of an angle of nip. Further, the score calculated for each of the spaces between the landmarks is added to a coordinate position and the coordinate position with the maximum score is estimated as the self-position.
Patent Document 2 discloses a method of estimating a self-position in consideration of the degree of coincidence when landmarks are detected. In the estimation method, the degree of coincidence between an object in a captured image and a landmark is calculated using a database that stores the positions of landmarks on a map and an amount of a characteristic. Thereafter, a self-existence range used to estimate the self-position in the order of the higher degree of coincidence of the landmark is determined. Accordingly, when a plurality of landmarks are extracted, a position error of a landmark with a low degree of coincidence can be reduced, and thus the self-position can be estimated with high accuracy.
Non-Patent Document 1 discloses a method of estimating a self-position by associating an image database with an input image. In the estimation method, a panorama image to which a position and a direction are assigned in advance is generated for each intersection. Characteristics extracted from the input image are associated with characteristics of a panorama image group and a vote is cast for a panorama image in which the association is possible. The position of a panorama image with the most vote results is determined as the self-position. Next, Homography between the input image and the panorama image is calculated, and the input image is projected on the panorama image using the calculated Homography, and the projection center is determined as a direction. In this case, a position at which the panorama image is captured is used as the position.